You probably have heard the statistic: One-quarter of Medicare spending is for patients in the last year of life. It is cited as a major reason for excessive medical spending in the US and leads to a widely-accepted conclusion: If only we would stop “wasting” dollars on futile care for those who soon will die anyway, we could significantly slow medical cost growth.
But what if the entire premise of this argument is wrong? What if Medicare spends just a small fraction of its resources on those who are expected to die? An important study concludes that only 5 percent of Medicare dollars—not 25 percent– is spent on individuals who were predicted to die. Thus, controlling end-of-life costs may do much less to reduce projected medical spending in the US than conventional wisdom claims.
A statistical fallacy
The research by Liran Einav, Amy Finkelstein, Sendhil Mullainathan, and Ziad Obermeyer provides important statistical support to buttress an argument that a handful of health policy experts such as Brookings Institution economist Henry J. Aaron have been making for years: Using spending in the last year of life as a proxy for futile care is deeply flawed—because we are very bad at predicting who is going to die and who is not.
As the authors write, the “common interpretations of end-of-life spending flirt with a statistical fallacy: Those who end up dying are not the same as those who were sure to die.”
Thus, much of the care we provide in what turns out to be the last year of life may not be futile at all, based on what doctors know at the time they are treating their patients. Many who are expected to die within, say, a year, don’t. And many who are not expected to die within that time frame do.
We are bad at predicting death
The article, published in the June edition of Science (paywall), is a dense statistical analysis of medical spending on patients the authors call the “ex post dead.” One of the authors, MIT professor Amy Finkelstein, just won a MacArthur genius award for her work in health care economics.
The authors use a technique called machine learning, where by analyzing massive amounts of data, computers can “learn” without being programmed. It is the same “big data” tool that makes it possible for amazon.com to recommend a future purchase based on your past buying habits.
To better understand our ability to predict death, they started by looked at millions of claims for Medicare enrollees. What they found was striking. Only 10 percent of those who had a 50 percent probability of passing away within a year in fact did die. Those considered most likely to die accounted for less than 5 percent of total spending, far less than the conventional estimate of 25 percent.
The problem: the cost of care for sick people
The authors found that it is difficult to accurately predict death even for those Medicare patients who have been hospitalized, or even those who have been hospitalized with metastatic cancer.
To go from big data to anecdotes, my wife, who is a hospice chaplain, often talks about how hard it is to predict when a patient will die. Although hospice enrollees, by definition, are expected to live six months or less, many live longer. Many who are thought to be within days of death live far beyond what medical professionals expect. And some, who hospice staff thinks will live for months, die unexpectedly within days. It is just hard to know.
For Einav, Finkelstein, and colleagues, the real story is not that we spend a lot of money on people at the end of life, it is that we spend a lot of money on people who are sick–whether they are dying or not. The authors acknowledge some patients who truly are dying do receive futile, high-cost care. But they do so far less often than the conventional wisdom suggests.
Their work should in no way be read to delegitimize the importance of comfort care or to discourage people from declining intensive medical treatment at the end of life. Those choices often are the best one’s for the well-being of the patients themselves.
But they are saying that whatever the benefits of palliative care versus aggressive treatment, major cost savings are not among them.
In short, we should be looking to how much we spend on sick people, not how much we spend on people who died within some arbitrary time frame. If we are going to cure the medical cost growth problem, we first need to accurately diagnose the problem. And Einav, Finkelstein and colleagues are telling us that, at least when it comes to one widely-believed cause of rising costs, we may have gotten it all wrong.